AI for Images and Genomes, Cancer Research at the Lunch Pitch

Image Processing for AI Made Easy

Pitch by Anders Garpebring, Associate professor, Department of Diagnostics and Intervention

The development and deployment of artificial intelligence (AI) models for image processing tasks require extensive preprocessing of image data to ensure it meets the requisite format and quality standards. This preparatory phase is critical, as it directly influences the performance and accuracy of AI models. However, it is often marked by complexity and labor-intensive procedures that can impede efficiency and scalability.

Anders Garpebring and collaborators developed Hero, an innovative graphical user interface designed to streamline the process of image data preparation and model deployment. During his pitch, Anders gave a demonstration of Hero, a user-friendly platform that allows users to construct pre-processing and post-processing pipelines through intuitive drawing of workflows, significantly reducing the technical barrier traditionally associated with these tasks. Hero also provides functionality for seamless deployment of trained models by employing a drag-and-drop mechanism, facilitating a smoother transition from model development to operational use. Finally, Hero is equipped to handle large batches of data, enabling the application of defined workflows on extensive datasets efficiently. This capability is particularly beneficial for projects requiring the processing of vast amounts of image data, ensuring both time and resource optimization.

With Hero, Anders hopes to simplify the preparatory stages of image data formatting and quality enhancement as well as accelerate the model deployment process. Through its graphical interface, he wants Hero to democratize access to advanced image processing techniques and become an invaluable tool for researchers, developers, and practitioners aiming to leverage AI in their work with minimal technical hurdles.
Anders is looking for researchers interested in using Hero and providing feedback, with the goal of improving or extending the software.

 
Exposure assessment during sensitive time periods – a key to understanding breast cancer risk?

Pitch by Sophia Harlid, Research Fellow, Department of Diagnostics and Intervention

Breast cancer etiology may be driven by disruptions and changes in breast tissue during critical windows of susceptibility such as puberty, menopause, and pregnancy. For example: during pregnancy, the breast tissue undergoes several critical stages of development and maturation, controlled by hormonal fluctuations. This makes the gland sensitive to environmental exposures, for example through contact with chemicals that interact with hormone receptors which could increase breast cancer risk. Effects triggered by such exposures could also serve as mediators between exposure and breast cancer development. In Sophia’s work she is investigating the effect of exposures during pregnancy to infer their effect on premenopausal breast cancer risk.
 
The new project involves register and biobank studies like the NorthPop study, as well as the setup of a new follow-up cohort (NorthMom) which will follow mothers from the NorthPop study from the time of their first breast cancer scan at age 40. Sophia is searching for potential collaborations that could help her setup a proof-of-concept study using animal models. She is also interested in any additional feedback or discussion about the NorthMom study. 

Deep-learning (DL) guided identification of LTR retrotransposon in plant genomes

Pitch by Jian-Feng Mao, Associate professor, Department of Plant Physiology, Umeå Plant Science Centre

Long Terminal Repeat retrotransposons (LTR-RTs) are a pervasive and dominant component in plant genomes, playing pivotal roles in functional variation, genome plasticity, and evolutionary processes. Despite their significance, the computational resources required for LTR-RT identification remain substantial, particularly for large genomes such as those of wheat (five to fifteen billion base pairs) and conifers (ten to forty billion base pairs). To address this, Jian-Feng and collaborators have developed a deep-learning (DL) guided method for LTR-RT identification, named LTR-checker. During his pitch, Jian-Feng showed the tool’s competitive precision, reduced computational demands, and enhanced flexibility. According to Jian-Feng, his work reveals the exciting potential of DL applications in genomics. The new method is publicly available at this link.

Jian-Feng’s aim is to unravel the complexity in plant genomes, by developing highly efficient computational tools. He is looking for collaborators from both plant biologists (to pinpoint critical research questions) and computer scientists (to integrate the latest advancements in computing technology).